An internal fault diagnosis method for oil-immersed transformer

Abstract

The application provides an oil-immersed transformer internal fault diagnosis method and relates to the technical field of power equipment fault diagnosis, correlation degree features are mined from time series data of the content of each characteristic gas dissolved in transformer oil, transformer fault diagnosis rules based on the correlation degree features are extracted, and logical expression of the diagnosis rules is realized in combination with fuzzy thought. The fault diagnosis rules are introduced into a transformer internal fault diagnosis model based on extreme learning machine as priori knowledge, so that the model has certain priori knowledge before data rules are mined, over-dependence of the model on samples is reduced, the model is corrected to some extent when the output result of the model conflicts with the priori knowledge, randomness caused by pure data driving is restricted, and the safety of the output result of the whole model is improved.

Description

Technical Field

[0001] This invention relates to the field of power equipment fault diagnosis technology, and more specifically, to a method for diagnosing internal faults in an oil-immersed transformer. Background Technology

[0002] Early fault diagnosis of oil-immersed transformers helps prevent fault escalation and improves maintenance efficiency. Dissolved Gas Analysis (DGA) has proven to be an effective method for diagnosing internal faults in oil-immersed power transformers worldwide. Thermodynamic studies have shown a correlation between transformer fault types and severity and the types, contents, and relative contents of characteristic gases, such as H2, CH4, C2H6, C2H4, C2H2, CO, and CO2. Through long-term engineering practice, engineers have gradually summarized certain knowledge in transformer fault diagnosis based on the results of Haast's thermodynamic research and accumulated experience, enabling preliminary judgments on the presence, type, and severity of transformer faults. Rule-based methods developed based on this knowledge include the Rogers four-ratio method, the IEC three-ratio method, the IEC research method, and the modified three-ratio method. These methods are simple, convenient, and based on a clear knowledge base, and have a wide range of applicability, achieving acceptable diagnostic results in most scenarios. However, it also has some drawbacks. First, the gas ratio boundary threshold is too absolute when coding. Second, the coding is incomplete, and in some cases, the fault type cannot be determined. Third, the coding is determined by human experience, but in fact, the gas production pattern during transformer faults is very complex, and the correspondence between characteristic gas content, relative content and fault type is difficult to obtain by induction. These problems all limit the diagnostic accuracy of the rule-based method based on domain knowledge in transformer fault diagnosis practice.

[0003] The widespread adoption of online dissolved gas analysis devices in transformer oil has made intelligent transformer diagnostic methods, based on statistical theory and driven by data, a research hotspot. By analyzing dissolved gases in oil and utilizing data mining methods such as machine learning, the mapping relationship between dissolved gases in transformer oil and internal fault types can be explored, thereby establishing a data-driven model for transformer fault diagnosis. Numerous studies have shown that, under the same dataset, data-driven intelligent diagnostic methods achieve higher diagnostic accuracy compared to traditional rule-based methods based on domain experience. However, these models suffer from several common problems: First, the technical level of online dissolved gas monitoring devices is still insufficient, and the characteristic gas data measured often contains significant errors, failing to meet the quality requirements of training data for data-driven models. Second, the low failure rate of power transformers and the varying probabilities of different fault types make it difficult to meet the requirements for fault sample size and sample balance among different fault types; the generalization performance of the model needs improvement when the fault sample size is insufficient. Third, data-driven models may exhibit serious diagnostic errors, threatening the safe operation of transformers. Summary of the Invention

[0004] To address the aforementioned shortcomings in the existing technology, the present invention provides a method for diagnosing internal faults in oil-immersed transformers, which solves the problem of insufficient accuracy in diagnostic results.

[0005] To achieve the above-mentioned objectives, the technical solution adopted by this invention is as follows:

[0006] A method for diagnosing internal faults in an oil-immersed transformer includes the following steps:

[0007] S1. Monitor oil-immersed transformers online and obtain a dataset with actual fault type labels;

[0008] S2. Based on the dataset, obtain the correlation characteristics, and through the fuzzy rule knowledge of internal transformer faults, diagnose the fault type and obtain the diagnosis result.

[0009] S3. Based on the dataset, the fault type diagnosis result is obtained by using the ELM (Extreme Learning Machine) model optimized by the PSO (Particle Swarm Optimization) algorithm.

[0010] S4. Based on the fault type diagnosis results of both steps S2 and S3, set a conflict function, combine it with the loss function, construct a knowledge function and use it as the fitness function of the ELM model, and optimize and train the ELM model through the PSO algorithm to obtain a data-knowledge joint-driven fault diagnosis model.

[0011] S5. Use data-knowledge joint-driven fault diagnosis model to diagnose internal faults in oil-immersed transformers.

[0012] Furthermore, the dataset includes: a first training set, a second training set, and a test set.

[0013] Further, step S1 includes the following sub-steps:

[0014] S11. Online monitoring of oil-immersed transformers to obtain time series data of the content of various characteristic gases dissolved in the oil;

[0015] S12. Organize the actual internal fault conditions of the transformer into actual fault type labels;

[0016] S13. The time series data of each characteristic gas content after being divided and labeled are the first training set, the second training set, and the test set.

[0017] Furthermore, the correlation feature is the value of various MIC coefficients (Maximal Information Coefficients) between the time series data of each characteristic gas content.

[0018] Furthermore, step S2 includes the following sub-steps:

[0019] S21. Calculate the values ​​of various MIC coefficients between the time series data of each characteristic gas content in the first and second training sets;

[0020] S22. Using the MIC coefficient as a binary classification indicator for fault type diagnosis, continuously adjust the classification threshold, use the first training set data, experiment and plot the ROC curve (receiver operating characteristic curve).

[0021] S23. Use the AUC (Area Under Curve) index of the ROC curve to select N kinds of MIC coefficients that have the best effect on binary classification of fault diagnosis, where N is a positive integer;

[0022] S24. Determine the classification thresholds for each of the N MIC coefficients using the Youden index of the ROC curve.

[0023] S25. By fuzzifying the classification thresholds of N kinds of MIC coefficients through membership functions, we can obtain fuzzy rule knowledge of internal faults in transformers.

[0024] S26. Based on the values ​​of N kinds of MIC coefficients in the second training set, the fault type diagnosis result is obtained through the fuzzy rule knowledge of internal transformer faults.

[0025] Furthermore, the membership function in step S25 includes an ascending semi-trapezoidal function and a descending semi-trapezoidal function. The ascending semi-trapezoidal function is as follows:

[0026]

[0027] Where f(m) is the ascending semi-trapezoidal function, m is the MIC coefficient value, a is the first slope parameter, b is the second slope parameter, and the value of b is equal to the classification threshold.

[0028] Furthermore, step S3 includes the following sub-steps:

[0029] S31. Set the basic parameters of the PSO algorithm and the ELM model, and randomly initialize the input layer weights and biases of the ELM model;

[0030] S32. Based on the first training set, use the PSO algorithm to train and optimize the ELM model;

[0031] S33. The ELM model trained in step S32 is used to diagnose faults on the second training set to obtain fault type diagnosis results.

[0032] Furthermore, step S4 includes the following sub-steps:

[0033] S41. Set a conflict function based on the fault type diagnosis results of both steps S26 and S33, and use cross-entropy as the loss function.

[0034] S42. Combine the conflict function and the loss function to construct a knowledge function;

[0035] S43. Use the knowledge function as the fitness function of the ELM model, optimize the training of the ELM model through the PSO algorithm to obtain the data-knowledge joint-driven fault diagnosis model, and verify the data-knowledge joint-driven fault diagnosis model through the test set.

[0036] Furthermore, the knowledge function is:

[0037]

[0038]

[0039] Where F(θ) is the knowledge function, θ is the parameter set of the ELM model, and x i Let h(θ,x) be the i-th sample data in the dataset. i ) represents the fault type diagnosis result obtained by diagnosing the i-th sample data through the ELM model, and t represents the fault type diagnosis result. i Let M be the actual label of the fault type for the i-th sample data, M be the number of samples, and β be the actual label of the fault type for the i-th sample data. iLet R(x) be the confidence level of the fuzzy rule knowledge about transformer internal faults for the i-th sample data. i P represents the fault type diagnosis result obtained from the i-th sample data through fuzzy rule knowledge diagnosis of transformer internal faults. i [( i ),h(θ,x i Let )] be the conflict function of the i-th sample, δ be the random factor, and rand i Let be the i-th random number, and α be the trust level value.

[0040] The beneficial effects of this invention are as follows:

[0041] (1) Correlation features were mined from the time series data of the content of various characteristic gases dissolved in transformer oil, and transformer fault diagnosis rules based on these correlation features were extracted. The logical representation of these diagnosis rules was realized by combining fuzzy logic. By introducing these fault diagnosis rules as prior knowledge into the transformer internal fault diagnosis model based on extreme learning machine, the model has certain prior knowledge before mining data patterns, thereby reducing its over-reliance on samples. At the same time, when the model output results conflict with the prior knowledge, the model is corrected to a certain extent, which restricts the arbitrariness that may be caused by pure data-driven approaches and improves the safety of the overall model output results.

[0042] (2) Based on the fuzzy rule knowledge of transformer fault diagnosis with correlation characteristics, the probability of serious diagnostic errors is greatly reduced by setting the conflict function.

[0043] (3) The fuzzy rule knowledge for transformer fault diagnosis based on correlation features has uncertainty when used for fault diagnosis, that is, its diagnostic results may not be accurate. This invention defines the concept of a confidence function to solve the negative impact that the introduction of uncertain knowledge may have on the model; defines the concept of a random factor to solve the problem of how to control the degree of introduction of uncertain knowledge; and establishes a knowledge function by combining the loss function and the conflict function, and uses the knowledge function as the target function for model training, so that the model training is subject to the dual constraints of existing prior knowledge and data. By comparing with the comparison model without prior knowledge and the current mainstream transformer internal fault diagnosis model, it is proved that this invention can effectively improve the generalization performance of the model; when the fault sample size is small, the fault diagnosis accuracy can be improved in this case because the fuzzy rule for transformer fault diagnosis based on correlation features is added as prior knowledge. Attached Figure Description

[0044] Figure 1 A flowchart of an internal fault diagnosis method for an oil-immersed transformer provided in an embodiment of the present invention;

[0045] Figure 2The membership function graph for MIC(C2H6:C2H4). Detailed Implementation

[0046] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0047] like Figure 1 As shown, in one embodiment of the present invention, a method for diagnosing internal faults in an oil-immersed transformer includes the following steps:

[0048] S1. Monitor oil-immersed transformers online and obtain a dataset with actual fault type labels. The dataset includes: a first training set, a second training set, and a test set.

[0049] Step S1 includes the following sub-steps:

[0050] S11. Monitor oil-immersed transformers online and obtain time series data of the content of various characteristic gases dissolved in the oil.

[0051] S12. Organize the actual fault conditions inside the transformer into actual fault type labels.

[0052] S13. The time series data of each characteristic gas content after being divided and labeled are the first training set, the second training set, and the test set.

[0053] In this embodiment, the online monitoring dataset includes time series data of the content of five characteristic gases: H2, CH4, C2H6, C2H4, and C2H2, along with their corresponding fault type labels. Transformer internal faults fall into three main categories: thermal faults, electrical faults, and mixed electrothermal faults. Based on the severity of the fault, they can be further divided into seven categories: low-temperature overheating (T1), medium-temperature overheating (T2), high-temperature overheating (T3), partial discharge (PD), low-energy discharge (D1), high-energy discharge (D2), and discharge combined with overheating faults (DT). Each fault type and its corresponding label are shown in Table 1.

[0054] Table 1 Transformer Fault Types and Their Corresponding Vector Labels

[0055]

[0056] S2. Based on the dataset, obtain the correlation characteristics and use the fuzzy rule knowledge of internal transformer faults to diagnose the fault type.

[0057] The correlation feature is the value of various MIC coefficients (Maximal Information Coefficients) between the time series data of each characteristic gas content.

[0058] Currently, DGA-based diagnostic methods primarily use the ratio of the content of various characteristic gases as fault diagnosis features. However, related research has demonstrated a correlation between the correlation between the time series data of characteristic gases and the fault type. The magnitude of this correlation could become a new fault diagnosis feature to distinguish different fault types in transformers. The requirement is no longer simply calculating the content ratio of two characteristic gases at a given moment, but rather exploring the correlation between the time series data of two sets of characteristic gases. Therefore, a method is needed to mine and quantitatively characterize the correlation between two variables from a large amount of data. MIC, proposed in 2014, is a method that can quantitatively characterize the correlation between two random variables and has proven to be a suitable tool for exploring correlations from large amounts of data. Its basic principles and calculation methods are not the focus of this patent and will not be explained here; please refer to relevant papers for details.

[0059] Step S2 includes the following sub-steps:

[0060] S21. Calculate the values ​​of various MIC coefficients between the time series data of each characteristic gas content in the first and second training sets.

[0061] This embodiment calculates the correlation between time series data of five commonly used characteristic gases, H2, CH4, C2H6, C2H4, and C2H2, based on MIC.

[0062] DL / T 722-2014 states that the total dissolved combustible gas content is also an important characteristic for detecting and identifying faults. Therefore, when calculating the correlation between a gas and other gases, the sum of the other four gases is also considered as a characteristic gas. For example, for H2, in addition to calculating the four values ​​MIC(H2:CH4), MIC(H2:C2H6), MIC(H2:C2H4), and MIC(H2:C2H2), the value of MIC(H2:SUM1) should also be calculated, where SUM1 is the sum of the concentrations of CH4, C2H6, C2H4, and C2H2. Finally, this embodiment requires the calculation of 15 sets of MIC values, as shown in Table 2.

[0063] Table 2. Correlation between characteristic gas time series data based on MIC.

[0064]

[0065] Note: SUM1=(CH4+C2H6+C2H4+C2H2); SUM2=(H2+C2H6+C2H4+C2H2); SUM3=(H2+CH4+C2H4+C2H2); SUM4=(H2+CH4+C2H6+C2H2); SUM5=(H2+CH4+C2H6+C2H4)

[0066] S22. Using the MIC coefficient as a binary classification indicator for fault type diagnosis, continuously adjust the classification threshold, use the first training set data, experiment and plot the ROC curve (receiver operating characteristic curve).

[0067] A set of MIC values ​​can be viewed as a binary classifier for classifying fault types during fault identification. Taking fault T2 as an example, in a binary classification problem, fault T2 is assigned a positive value, while all other fault types are uniformly assigned a negative value. For a certain correlation feature MIC... i Set a threshold P, when MIC i When the value is higher than P, the fault is judged as a T2 fault; otherwise, it belongs to one of the other fault types. Various fault types can be classified by combining multiple binary classifiers, but features with poor classification performance will actually reduce the final diagnostic accuracy. Therefore, it is necessary to study the classification performance and classification threshold of each binary classifier.

[0068] ROC curves are suitable for binary classification problems and are mainly used to determine the threshold of a classifier and evaluate its performance. In ROC analysis, there are four classification results: TP (True Positive), where both the actual fault and the classification result are T2 faults; FP (False Positive), where the actual fault is not T2, but the classification result is T2; TN (True Negative), where neither the actual fault nor the classification result is T2 fault; and FN (False Negative), where the actual fault is T2, but the classification result is not T2.

[0069] The false positive rate is defined as R_FP:

[0070]

[0071] The true positive rate is defined as R_TP:

[0072]

[0073] In the above formula, N_FP, N_TN, B_TP, and N_FN are the number of samples classified as FP, TN, TP, and TN, respectively.

[0074] Using R_FP as the x-axis and R_TP as the y-axis, by continuously adjusting the threshold P, the same sample data will produce different points in a two-dimensional plane graph. Connecting these points forms the ROC curve.

[0075] S23. Use the AUC (Area Under Curve) index of the ROC curve to select N MIC coefficients that have the best effect on binary classification of fault diagnosis, where N is a positive integer.

[0076] The area under the curve (ROC) and the horizontal axis is called the AUC (Area Under Curve), which ranges from 0 to 1. A larger AUC value indicates better classification performance based on that feature. Therefore, by comparing the AUC values ​​of different correlation features under the same fault type, their performance in classifying that type of fault can be determined. Generally, an AUC value above 0.75 indicates that the feature has a good classification effect. This embodiment, by comparing the AUC values, identified six MIC features with the best classification performance for T2 faults: MIC(C2H6:SUM3), MIC(C2H6:C2H2), MIC(H2:SUM1), MIC(C2H2:SUM5), MIC(C2H4:C2H2), and MIC(CH4:C2H2). Combining these features can form diagnostic rules for T2 faults.

[0077] S24. Determine the classification thresholds for each of the N MIC coefficients using the Youden index of the ROC curve.

[0078] The Youden index is equal to the ordinate of a point on the ROC curve minus its abscissa. The optimal classification threshold is the one where the Youden index is maximized at each point.

[0079] The fuzzy rules for transformer fault diagnosis based on correlation characteristics use thermodynamic analysis as a knowledge base. They have the advantages of good adaptability in a wide range of fields and are not prone to serious diagnostic errors. They are simple and practical. The specific fault classification rules are shown in Table 3.

[0080] Table 3 Fault characteristics and classification thresholds for different fault types

[0081]

[0082]

[0083] However, this method of diagnosis using feature thresholds suffers from overly defined boundaries. Fuzzy reasoning is widely used to solve uncertainty problems, and this embodiment introduces it into transformer fault diagnosis based on correlation features, extracting fuzzy rule knowledge for diagnosing various fault types.

[0084] S25. By fuzzifying the classification thresholds of the N types of MIC coefficients using membership functions, we can obtain fuzzy rule knowledge of internal faults in transformers.

[0085] Membership functions include ascending semi-trapezoidal functions and descending semi-trapezoidal functions. The ascending semi-trapezoidal function is:

[0086]

[0087] Where f(m) is the ascending semi-trapezoidal function, m is the MIC coefficient value, a is the first slope parameter, b is the second slope parameter, and the value of b is equal to the classification threshold.

[0088] The descending half-trapezoidal function is similar. For the descending half-trapezoidal function, the value of 'a' is equal to the classification threshold. In this embodiment, the difference between 'a' and 'b' is set to 0.14, thus making the slope a fixed value.

[0089] From the threshold values ​​of each feature shown in Table 3, it can be seen that if MIC(C2H6:C2H4) < 0.48, the fault type may be D1 fault; if MIC(C2H6:C2H4) > 0.71, the fault type may be T3 fault. Therefore, for the ascending half-trapezoidal membership function, b is set to 0.71 and a is set to 0.57 (0.71-0.14); for the descending half-trapezoidal membership function, a is set to 0.48 and b is set to 0.62 (0.48+0.14). When the calculated MIC value is greater than zero and less than 0.57, it is determined to be in state I; when the calculated MIC value is greater than 0.62, it is determined to be in state II; when the calculated MIC value is between 0.57 and 0.62, it is determined to be in both state I and II. Therefore, the membership function of MIC(C2H6:C2H4) is as follows: Figure 2 As shown.

[0090] After fuzzifying each feature quantity according to Table 3, the resulting fuzzy rule knowledge of transformer internal faults is shown in Table 4.

[0091] S26. Based on the values ​​of N kinds of MIC coefficients in the second training set, the fault type diagnosis result is obtained through the fuzzy rule knowledge of internal transformer faults.

[0092] Table 4. Transformer Fuzzy Diagnosis Rules Based on Correlation Features

[0093]

[0094] S3. Based on the dataset, the fault type diagnosis result is obtained by using the ELM (Extreme Learning Machine) model optimized by the PSO (Particle Swarm Optimization) algorithm.

[0095] Step S3 includes the following sub-steps:

[0096] S31. Set the basic parameters of the PSO algorithm, including population size, learning factor, and number of iterations. Set the basic parameters of the ELM model, including the number of nodes in the input layer, hidden layer, and output layer, and the activation function of the hidden layer. Randomly initialize the input layer weights and biases of the ELM model.

[0097] In this embodiment, the population size of the PSO optimization algorithm is set to 40, both learning factors are set to 1.5, and the number of iterations is set to 1000.

[0098] The ratio of H2 to the total content of all characteristic gases, and the ratios of CH4, C2H6, C2H4, and C2H2 to the total content of hydrocarbon gases, are used as inputs to the ELM model. Therefore, the number of neurons in the input layer is set to 5. Since the input values ​​are all between 0 and 1, there is no issue of magnitude differences affecting fault diagnosis, and no normalization processing is required for the input data. The vector labels for the seven fault types of transformers are defined as shown in Table 1. The number of neurons in the output layer is set to 7, and the activation function of the hidden layer neurons is the Sigmoid function, with 11 neurons in the hidden layer. For the output vector, the Softmax function is used to process it so that the value of each output element in the output vector is between 0 and 1. At this point, its value represents the probability value of the input data belonging to each fault type. If the output vector corresponding to a certain input data is (0.015, 0.322, 0.458, 0.014, 0.024, 0.064, 0.103), then the seven elements correspond to the probabilities of belonging to the seven fault types T1, T2, T3, PD, D1, D2, and DT, respectively. The ELM model selects the fault type corresponding to the maximum value as the fault type corresponding to the data set. For this data set, it is classified as a T3 fault.

[0099] S32. Based on the first training set, use the PSO algorithm to train and optimize the ELM model.

[0100] S33. The ELM model trained in step S32 is used to diagnose faults on the second training set to obtain fault type diagnosis results.

[0101] S4. Based on the fault type diagnosis results of steps S2 and S3, set a conflict function, combine it with the loss function, construct a knowledge function and use it as the fitness function of the ELM model, and optimize and train the ELM model through the PSO algorithm to obtain a data-knowledge joint-driven fault diagnosis model.

[0102] Step S4 includes the following sub-steps:

[0103] S41. Set a conflict function according to the fault type diagnosis results of both step S26 and step S33, and use cross-entropy as the loss function.

[0104] In this embodiment, according to the actual situation in the field of oil-immersed transformer fault diagnosis, two types of serious diagnostic errors of the pure data-driven model are defined as follows: One is the diagnostic error between major fault categories, such as diagnosing an electrical fault as a thermal fault; the other is diagnosing a serious fault as a non-serious fault, such as diagnosing high-temperature overheating as low-temperature overheating. Based on the above definitions, the calculation rules for setting the conflict function are as follows:

[0105] Rule 1: If (A <= 3 and B > 3) and (3 < A <= 6 and (B <= 3 or B = 7)) and (A = 7 and B < 7), then the conflict function value is incremented by 1.

[0106] Rule 2: If (A <= 3 and B <= 3 and A < B) and (A >= 4 and A <= 6 and B >= 4 and B <= 6 and A < B), then the conflict function value is incremented by 1.

[0107] Where A is the column number where the maximum value in the output vector of the ELM model is located. Combining the settings in Table 1, it can be seen that A represents the probability that the pure data-driven model believes a set of input data belongs to the fault corresponding to this column is the largest. B is the column number where the maximum value in the diagnostic result obtained from the fuzzy rules knowledge of transformer fault diagnosis based on correlation characteristics is located, representing the fault type that the fuzzy rules knowledge of transformer fault diagnosis based on correlation characteristics believes a set of input data belongs to.

[0108] The meaning of Rule 1 is that when the fault types are classified into three major categories: thermal faults, electrical faults, and electro-thermal hybrid faults, if the output result of the ELM model and the diagnostic result obtained from the fuzzy rules knowledge are not in the same major category, the conflict function value is increased to impose guidance on the training of the ELM model and reduce the probability of its occurrence of low-level diagnostic errors. Of course, the result obtained from the fuzzy rules knowledge is not always correct. To prevent its error from having a negative impact on the training of the ELM model, this patent will solve this problem by introducing a random factor and a confidence function later. The meaning of Rule 2 is that if the output result of the ELM model and the diagnostic result obtained from the fuzzy rules knowledge are in the same major category of faults, but their diagnoses of the fault severity are different, and the fault severity considered by the ELM model is lower than that considered by the knowledge rules, the conflict function value is also increased, thereby reducing the probability of diagnosing a serious fault as a non-serious fault and ensuring the safety of the output result.

[0109] S42. Combine the conflict function and the loss function to construct a knowledge function.

[0110] S43. Using the knowledge function as the fitness function of the ELM model, the PSO algorithm is used to optimize the input layer weights and biases of the ELM model. In each iteration, the current individual particle optimal value and the global optimal value are determined, and the particle velocity and position are updated accordingly. After reaching the convergence condition, the optimal input layer weights and biases of the ELM model are output to complete its optimization, and the final data-knowledge joint-driven fault diagnosis model is obtained. The data-knowledge joint-driven fault diagnosis model is then validated through the test set.

[0111] The knowledge function is:

[0112]

[0113]

[0114] Where F(θ) is the knowledge function, θ is the parameter set of the ELM model, and x i Let h(θ,x) be the i-th sample data in the dataset. i ) represents the fault type diagnosis result obtained by diagnosing the i-th sample data through the ELM model, and t represents the fault type diagnosis result. i Let H[h(θ,x) be the actual label of the fault type for the i-th sample data. i ),t i [ ] represents the cross-entropy between the fault type diagnosis result obtained from the ELM model for the i-th sample data and the actual fault type label, where M is the number of samples and β is the cross-entropy between the two samples. i Let R(x) be the confidence level of the fuzzy rule knowledge about transformer internal faults for the i-th sample data. i P represents the fault type diagnosis result obtained from the i-th sample data through fuzzy rule knowledge diagnosis of transformer internal faults. i [( i ),h(θ,x i Let )] be the conflict function of the i-th sample, δ be the random factor, and rand i Let be the i-th random number, and α be the trust level value.

[0115] Using the cross-entropy function as the loss function, the knowledge function can be obtained by adding the calculated cross-entropy function and the conflict function. The cross-entropy function reflects the error between the data mining algorithm output and the actual label, while the conflict function reflects the conflict between the data mining algorithm and the fuzzy rule knowledge of transformer fault diagnosis based on correlation features.

[0116] Considering that diagnostic methods based on fuzzy rule knowledge are not always correct, this embodiment draws on the greedy algorithm in reinforcement learning, introduces the concept of random factors into the knowledge function, and sets the concept of trust level to overcome the guidance bias caused by the diagnostic errors of the fuzzy rule knowledge-based method on the training of the ELM model.

[0117] The random factor δ reflects the degree of involvement of fuzzy rule knowledge for transformer fault diagnosis based on correlation features in model training. Its value is related to the size of the sample data used in model training. If the sample size is sufficient, the purely data-driven model can perform well, and in this case, a larger value should be taken to reduce the excessive involvement of fuzzy rule knowledge for transformer fault diagnosis based on correlation features. Conversely, when the sample size is insufficient, a smaller value should be taken to fully utilize the guiding role of fuzzy rule knowledge for transformer fault diagnosis based on correlation features, providing some prior knowledge for the overall model, reducing excessive dependence on sample size, and improving the overall generalization performance of the model. The value of α represents the degree of trust in domain experience knowledge. When applying deterministic knowledge, its value can be larger. When the introduced domain knowledge is uncertain, the value of α should be reasonably adjusted according to the uncertainty of the knowledge. Furthermore, considering that as the data-knowledge jointly driven fault diagnosis model iterates through training, the ELM model output will gradually become more accurate, the value of α should show a gradually decreasing trend; that is, the value of α is not a fixed value but should change with the number of iterations.

[0118] S5. Use data-knowledge joint-driven fault diagnosis model to diagnose internal faults in oil-immersed transformers.

[0119] To verify the effectiveness of this embodiment, the data-knowledge joint-driven diagnostic model was compared with the extreme learning machine model without the addition of fuzzy rule knowledge for transformer fault diagnosis based on correlation features. The fault types and their sample sizes in the training set 2 and the test set are divided according to Table 5. The diagnostic accuracy of the two models on the test set is shown in Table 6. The distribution of misdiagnosis on the test set and the number of serious faults diagnosed as non-serious faults are shown in Tables 7 and 8, respectively.

[0120] Table 5. Fault types and sample sizes in the training and test sets.

[0121]

[0122] Table 6. Diagnostic accuracy of the two models on the test set.

[0123]

[0124] Table 7. Misdiagnosis Distribution of the Original Model and the Data-Knowledge Jointly Driven Model

[0125]

[0126] Table 8. Number of samples diagnosed as non-serious faults in severe fault diagnosis.

[0127]

[0128] In summary, this invention incorporates fuzzy rule knowledge for transformer fault diagnosis based on correlation features into the training process of the ELM model, using the degree of conflict between this knowledge and the data-driven model results as one of the indicators for evaluating the model's training effectiveness. Based on the actual needs of transformer fault diagnosis, two calculation rules for the conflict function are defined to limit the model from making serious diagnostic errors, i.e., diagnosing serious faults as non-serious faults or making diagnostic errors between fault categories. Comparison with a purely data-driven transformer fault diagnosis model shows that this invention can significantly reduce the probability of serious diagnostic errors. However, the fuzzy rule knowledge for transformer fault diagnosis based on correlation features has uncertainty when used for fault diagnosis; that is, its diagnostic results may not be accurate. This invention addresses this issue. The concept of a trust function is defined to address the potential negative impact of introducing uncertain knowledge on the model. The concept of a random factor is also defined to address the issue of controlling the degree of introduction of uncertain knowledge. A knowledge function is established by combining the loss function and the conflict function, and this knowledge function is used as the objective function for model training. This makes the model training subject to the dual constraints of existing prior knowledge and data. Comparison with a model without prior knowledge and current mainstream transformer internal fault diagnosis models demonstrates that this invention can effectively improve the model's generalization performance. When the fault sample size is small, the addition of fuzzy rules for transformer fault diagnosis based on correlation features as prior knowledge improves the fault diagnosis accuracy in this case.

[0129] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An oil-immersed transformer internal fault diagnosis method characterized by comprising: Includes the following steps: S1. Monitor oil-immersed transformers online and obtain a dataset with actual fault type labels; S2. Based on the dataset, obtain the correlation characteristics, and through the fuzzy rule knowledge of internal transformer faults, diagnose the fault type and obtain the diagnosis result. S3. Based on the dataset, the fault type diagnosis results are obtained through the ELM model optimized by the PSO algorithm. S4. Based on the fault type diagnosis results of both steps S2 and S3, set a conflict function, combine it with the loss function, construct a knowledge function and use it as the fitness function of the ELM model, and optimize and train the ELM model through the PSO algorithm to obtain a data-knowledge joint-driven fault diagnosis model. S5. Diagnose internal faults in oil-immersed transformers using a data-knowledge joint-driven fault diagnosis model; The dataset includes: a first training set, a second training set, and a test set; Step S1 includes the following sub-steps: S11. Online monitoring of oil-immersed transformers to obtain time series data of the content of various characteristic gases dissolved in the oil; S12. Organize the actual internal fault conditions of the transformer into actual fault type labels; S13. The time series data of each characteristic gas content after being divided and labeled are the first training set, the second training set, and the test set; The correlation feature is the value of various MIC coefficients between the time series data of each characteristic gas content; Step S2 includes the following sub-steps: S21. Calculate the values ​​of various MIC coefficients between the time series data of each characteristic gas content in the first and second training sets; S22. Using the MIC coefficient as a binary classification indicator for fault type diagnosis, continuously adjust the classification threshold, use the first training set data, experiment and plot the ROC curve. S23, select the best binary classification effect of fault diagnosis by using the AUC index of the ROC curve N A MIC coefficient, N is a positive integer; S24, determining by the Youden index of the ROC curve N a respective classification threshold for each MIC coefficient; S25, obtaining the fuzzy rule knowledge of the transformer internal fault by fuzzy processing the respective classification threshold of the MIC coefficient. N A MIC coefficient is obtained, and the respective classification threshold of the MIC coefficient is fuzzy processed to obtain fuzzy rule knowledge of a transformer internal fault. S26、According to the second training set N A value of the MIC coefficient is obtained by diagnosing the fault type diagnosis result through the transformer internal fault fuzzy rule knowledge. Step S3 includes the following sub-steps: S31. Set the basic parameters of the PSO algorithm and the ELM model, and randomly initialize the input layer weights and biases of the ELM model; S32. Based on the first training set, use the PSO algorithm to train and optimize the ELM model; S33. The ELM model trained in step S32 is used to diagnose faults on the second training set to obtain fault type diagnosis results. Step S4 includes the following sub-steps: S41. Set a conflict function based on the fault type diagnosis results of both steps S26 and S33, and use cross-entropy as the loss function. S42. Combine the conflict function and the loss function to construct a knowledge function; S43. Use the knowledge function as the fitness function of the ELM model, optimize the training of the ELM model through the PSO algorithm to obtain the data-knowledge joint-driven fault diagnosis model, and verify the data-knowledge joint-driven fault diagnosis model through the test set. The knowledge function is: in, For knowledge functions, For the parameter set of the ELM model, For the first in the dataset One sample data, For the first Fault type diagnosis results obtained from individual sample data using the ELM model For the first The actual label of the fault type for each sample data. For the first The cross-entropy between the fault type diagnosis results obtained from the ELM model and the actual fault type label of a sample data. For the sample size, For the first The reliability of fuzzy rule knowledge regarding transformer internal faults based on a sample of data. For the first Fault type diagnosis results obtained from sample data using fuzzy rule knowledge diagnosis of transformer internal faults For the first The conflict function of each sample As a random factor, For the first A random number, This represents the level of trust.

2. The method for diagnosing internal faults in an oil-immersed transformer according to claim 1, characterized in that, The membership function in step S25 includes an ascending semi-trapezoidal function and a descending semi-trapezoidal function. The ascending semi-trapezoidal function is: in, For ascending semi-trapezoidal functions, This is the MIC coefficient value. The first slope parameter, This is the second slope parameter. The value of is equal to the classification threshold.

Images